Minima order condition

[araihC]

Minima method to trigger detection update. Transition time are identified as the deepest minima candidate between two good maxima.

• Local maxima are identified as maxima provided that:
  1. Two maxima are separated by at least MIN_PEAK_DISTANCE_MAXIMA
  2. The intensity is higher than MAXIMA_THRESHOLD_FRACTION than the amplitude span
• Local minima are identified as transition candidates provided that the signal in the following MINIMA_PERSISTENCE is higher. To implement this the scipy argrelmin funcion has been slightly modified to set the minima order parameter only on one side of the signal acquisition.

[rgutzen]

Looking at the figure above, it seems that the selected minima are just the first before a selected maxima (e.g 12th minima, or 9th from last in figure above) which would be a much easier algorithm. Also could you explain the purpose of setting the minima order parameter only to one side?

[araihC]

Yes, this is how the result should look like: this algorithm should be used on smooth signals (at which spatial resolution are these recordings measured? maybe a spatial smoothing should be applied). Indeed, we aim at collecting all transition candidates (e.g. the transition of two close consecutive waves) to be further cleaned up by the wave hunt method in stage 04. We thus prefer to identify a "wrong" up transition candidate rather then not collecting a "right" one.

Looking at the figure above, it seems that the selected minima are just the first before a selected maxima (e.g 12th minima, or 9th from last in figure above) which would be a much easier algorithm.

Yes, this is an error in my comment, the last minima before a transition is selected since optical signals typically have a fast response when moving from down to up states

Also could you explain the purpose of setting the minima order parameter only to one side?

Due to the fast response from down to up state, we expect an approximately flat behavior in the down state (not well represented by a parabolic branch) rapidly changing to the up-state, modeled as a parabolic branch.

Hope this clears up your doubts.

[rgutzen]

We thus prefer to identify a "wrong" up transition candidate rather then not collecting a "right" one.

I understand that you want to implement a method that is less strict in the minima selection. However, I do not see that this is the case here as, for example, in the above plot there are already several minima that I would consider "right" and that are not collected.

this algorithm should be used on smooth signals (at which spatial resolution are these recordings measured? maybe a spatial smoothing should be applied). Indeed, we aim at collecting all transition candidates (e.g. the transition of two close consecutive waves) to be further cleaned up by the wave hunt method in stage 04.

Of course, different blocks may work better or worse in certain combinations, but a key aspect of the pipeline design is the modularity, which also means that blocks should be as independent of other blocks and stages as possible, i.e. this method shouldn't strongly rely on any particular filtering before or after, in order to be reasonable or useful.

Therefore, I believe that we can make the detection criteria and the implementation more precise. We can maybe discuss this also in another meeting.

[araihC]

Updated Minima transition algorithm. Now it works like this:

• Local maxima are required to be at least 280ms distant (otherwise the higher is selected) and higher than a set threshold [red points]
• the threshold is dynamically detected as 0.4*span of the signal within a 4s time window [black line]
• local minima candidates are required to be at leas 280ms distant (otherwise the lower is selected) and with a persistence of 160ms (next 160ms needs to be higher). [orange points]
• between two maxima only one minimum is accepted: the last one is the selected [green points]

The persistence algorithm needs to be asymmetric (only on one side) to be take into consideration the possibility of flat down-states. This is what the output should look like

I also added a plot of a representative pixel(s) in the pipeline.

[rgutzen]

I still see some improvements we can apply to this method, but I think it can also be handled in the master branch as it is running and not interfering with the rest of the pipeline architecture.